See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/221346493

An Evaluation on Feature Selection for Text Clustering

Conference Paper · January 2003

Source: DBLP

CITATIONS READS
222 274

4 authors, including:

Wei-Ying Ma
Toutiao
389 PUBLICATIONS   28,256 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Wei-Ying Ma on 03 June 2014.

The user has requested enhancement of the downloaded file.

 An Evaluation on Feature Selection for Text Clustering

Tao Liu LTMAILBOX@263.SINA.COM

Department of Information Science, Nankai University, Tianjin 300071, P. R. China

Shengping Liu LSP@IS.PKU.EDU.CN

Department of Information Science, Peking University, Beijing 100871, P. R. China

Zheng Chen ZHENGC@MICROSOFT.COM

Wei-Ying Ma WYMA@MICROSOFT.COM
Microsoft Research Asia, 49 Zhichun Road, Beijing 100080, P. R. China

Abstract the inherent data sparsity. Obviously, a single document has

Feature selection methods have been successfully
applied to text categorization but seldom applied
to text clustering due to the unavailability of class
label information. In this paper, we first give
empirical evidence that feature selection methods
can improve the efficiency and performance of text
clustering algorithm. Then we propose a new
feature selection method called “Term
Contribution (TC)” and perform a comparative
study on a variety of feature selection methods for
text clustering, including Document Frequency
(DF), Term Strength (TS), Entropy-based (En),
Information Gain (IG) and ‫א‬2 statistic (CHI).
Finally, we propose an “Iterative Feature Selection
(IF)” method that addresses the unavailability of
label problem by utilizing effective supervised
feature selection method to iteratively select
features and perform clustering. Detailed
experimental results on Web Directory data are
provided in the paper.
1. Introduction
Text clustering is one of the central problems in text mining
and information retrieval area. The task of text clustering is
to group similar documents together. It had been applied to
several applications, including improving retrieval
efficiency of information retrieval systems (Kowalski,
1997), organizing the results returned by a search engine in
response to user’s query (Zamir et al., 1997), browsing
large document collections (Cutting et al., 1992), and
generating taxonomy of web documents (Koller & Sahami,
1997), etc.
In text clustering, a text or document is always represented
as a bag of words. This representation raises one severe
problem: the high dimensionality of the feature space and
a sparse vector over the set of all terms. The performance of
clustering algorithms will decline dramatically due to the
problems of high dimensionality and data sparseness
(Aggrawal &Yu, 2000). Therefore it is highly desirable to
reduce the feature space dimensionality. There are two
commonly used techniques to deal with this problem:
feature extraction and feature selection. Feature extraction
is a process that extracts a set of new features from the
original features through some functional mapping (Wyse et
al., 1980), such as principal component analysis (PCA)
(Jolliffe, 1986) and word clustering (Slonim & Tishby,
2000). The feature extraction methods have a drawback that
the generated new features may not have a clear physical
meaning so that the clustering results are difficult to
interpret (Dash & Liu, 2000).
Feature selection is a process that chooses a subset from the
original feature set according to some criterions. The
selected feature retains original physical meaning and
provides a better understanding for the data and learning
process. Depending on if the class label information is
required, feature selection can be either unsupervised or
supervised. For supervised methods, the correlation of each
feature with the class label is computed by distance,
information dependence, or consistency measures (Dash &
Liu, 1997). Further theoretical study based on information
theory can be found on (Koller & Sahami, 1996) and
complete reviews can be found on (Blum & Langley, 1997;
Jain et al., 2000; Yang & Pedersen, 1997).
As for feature selection for clustering, there have some
works on it. Firstly, any traditional feature selection method
that does not need the class information, such as document
frequency (DF) and term strength (TS) (Yang, 1995), can
be easily applied to clustering. Secondly, there are some
newly proposed methods, for example, entropy-based
feature ranking method (En) is proposed by Dash and Liu
(2000) in which feature importance is measured by the

Proceedings of the Twentieth International Conference on Machine Learning (ICML-2003), Washington DC, 2003.
contribution to an entropy index based on the data
similarity; the individual “feature saliency” is estimated and
an Expectation-Maximization (EM) algorithm using
Minimum Message Length criteria is derived to select the
feature subset and the number of clusters (Martin et al.,
2002).
While the methods mentioned above are not directly
targeted to clustering text documents, in this paper we
introduce two novel feature selection methods for text
clustering. One is Term Contribution (TC) which ranks the
feature by its overall contribution to the documents
similarity in a dataset. Another is Iterative Feature Selection
(IF), which utilizes some successful feature selection
methods (Yang & Pedersen, 1997), such as Information
Gain (IG) and ‫א‬2 statistics (CHI), to iteratively select
features and perform text clustering at the same time.
Another contribution of this paper is a comparative study
on feature selection for text clustering. We investigate (a) to
what extent feature selection can improve the clustering
quality, (b) how much of the document vocabulary can be
reduced without losing useful information in text clustering,
(c) what the strengths and weaknesses of existing feature
selection methods are when applied to text clustering, and
(d) what the difference is among the results of different
datasets. In this paper, we try to address these problems by
empirical evidences. We first show that feature selection
methods can improve the efficiency and performance of text
clustering in ideal cases, in which the class label for each
document is already known. Then we perform a
comparative study on various feature selection methods for
text clustering. Finally, we evaluate the performance of
iterative feature selection method based on K-means using
entropy and precision measures.
The rest of this paper is organized as follows. In Section 2,
we give a brief introduction on several feature selection
methods and propose a new feature selection method, called
Term Contribution. In Section 3, we propose a new iterative
feature selection method that utilizes the supervised feature
selection algorithm without the need to know the class
information in advance. In Section 4, we conduct several
experiments to compare the effectiveness of different
feature selection methods in ideal and real cases. Finally,
we summarize our major contributions in Section 5.
2. Feature Selection Methods
In this Section, we give a brief introduction on several
effective feature selection methods, including two
supervised methods, IG and CHI, and four unsupervised
methods, DF, TS, En and TC. All these methods assign a
score to each individual feature and then select features
which are greater than a pre-defined threshold.
In the following, let D denote the documents set, M the
dimension of the features, and N the number of documents
in the dataset.
2.1 Information Gain (IG)
Information gain (Yang & Pedersen, 1997) of a term
measures the number of bits of information obtained for
category prediction by the presence or absence of the term
in a document. Let m be the number of classes. The
information gain of a term t is defined as
IG(t ) = −∑i =1 p(ci ) log p(ci )
m
+ p(t )∑i =1 p(ci | t ) log p (c i | t )
m (1)
+ p(t )∑i =1 p(ci | t ) log p (c i | t )
m
2.2 ‫א‬2 statistic (CHI)
The ‫א‬2 statistic measures the association between the term
and the category (Galavotti et al., 2000). It is defined to be
N × ( p(t , c) × p (t ,c) − p(t , c) × p(t , c)) 2 (2)
χ 2 (t , c) =
p(t ) × p(t ) × p(c) × p(c)
m
(3)
χ 2 (t ) = avg{χ 2 (t ,c i )}
i =1
2.3 Document Frequency (DF)
Document frequency is the number of documents in which a
term occurs in a dataset. It is the simplest criterion for term
selection and easily scales to a large dataset with linear
computation complexity. It is a simple but effective feature
selection method for text categorization (Yang & Pedersen,
1997).
2.4 Term Strength (TS)
Term strength is originally proposed and evaluated for
vocabulary reduction in text retrieval (Wilbur & Sirotkin,
1992), and later applied to text categorization (Yang, 1995).
It is computed based on the conditional probability that a
term occurs in the second half of a pair of related
documents given that it occurs in the first half:
TS (t ) = p (t ∈ d j | t ∈ d i ), d i , d j ∈ D I sim(d i , d j ) > β (4)
where β is the parameter to determine the related pairs.
Since we need to calculate the similarity for each document
pair, the time complexity of TS is quadratic to the number
of documents. Because the class label information is not
required, this method is also suitable for term reduction in
text clustering.
2.5 Entropy-based Ranking (En)
Entropy-based ranking is proposed by Dash and Liu (2000).
In this method, the term is measured by the entropy
reduction when it is removed. The entropy is defined as the
equation (5):
 E (t ) = −∑ ∑ ( S i , j × log(S i , j ) + (1 − S i , j ) × log(1 − S i , j )), (5)
N N
i =1 j =1
where Si , j is the similarity value between the document d i
and d j . Si , j is defined as the equation (6):
−α × disti , j ln(0.5)
S i, j = e ,α = −
dist
where dist i , j is the distance between the document di and
d j after the term t is removed, dist is the average distance
among the documents after the term t is removed.
The most serious problem of this method is its high
computation complexity O ( MN 2 ) . It is impractical when
there is a large number of documents and terms, and
therefore, sampling technique is used in real experiments
(Dash & Liu, 2000).
2.6 Term Contribution (TC)
We introduce a new feature selection method called “Term
Contribution” that takes the term weight into account.
Because the simple method like DF assumes that each term
is of same importance in different documents, it is easily
biased by those common terms which have high document
frequency but uniform distribution over different classes.
TC is proposed to deal with this problem,
The result of text clustering is highly dependent on the
documents similarity. So the contribution of a term can be
viewed as its contribution to the documents’ similarity. The
similarity between documents d i and d j is computed by
dot product:
sim ( d i , d j ) = ∑ f (t , d i ) × f (t , d j )
t
where f (t , d ) represents the tf*idf (Salton, 1989) weight of
term t in document d .
So we define the contribution of a term in a dataset as its
overall contribution to the documents’ similarities. The
equation is
∑ f (t , d ) × f (t , d
TC (t ) =
i j
i , j Ii ≠ j
“ltc” scheme is used to compute each term’s tf*idf value
which takes the log of term’s frequency in the document,
then multiplies it by the IDF weight of this term, and finally
normalizes the document length.
If the weights of all terms are equal, we simply set
f (t , d ) = 1 when the term t appears in the document d .
Then the value TC(t) can be written as the equation (9):
TC (t ) = DF (t )( DF (t ) − 1)
Since the transformation is increasing monotonously while
the DF(t) (the document frequency of the term t) is a
positive integer, DF is just the special case of TC.
Using the inverse document indexing technology (Salton,
1989), the time complexity of TC is O( MN 2 ) , where N is
the average documents per term occurs in.
3. Iterative Feature Selection (IF) Method
(6)
Feature selection methods have successfully applied to text
categorization for long years. Feature selection can
dramatically improve the efficiency of text categorization
algorithm by removing up to 98% unique terms and even
improve the categorization accuracy to some extent (Yang
& Pedersen, 1997). So it is an interesting idea to apply
feature selection methods to text clustering task to improve
the clustering performance. In order to test our idea, several
experiments were conducted. As can be seen from Section
4.3.1 and Section 4.3.2, if the class label information is
known, supervised feature selection methods, such as CHI
and IG, are much more efficient than unsupervised feature
selection methods for text clustering, not only more terms
can be removed, but also better performance can be yielded.
Supervised feature selection method can not be directly
applied to text clustering because of the unavailability of
the required class label information. Fortunately, we found
that clustering performance and feature selection can be
reinforced by each other. On the one hand, good clustering
results will provide good class labels to select better
features for each category; on the other hand, better features
will help clustering to improve the performance to provide
better class labels. Enlightened by the EM algorithm, we
propose a novel iterative feature selection method, which
utilizes supervised feature selection methods for text
clustering methods.
(7)
EM is a class of iterative algorithms for maximum
likelihood estimation in problems with incomplete data
(Dempster et al., 1977). Suppose we have a dataset
generated from a model with parameters and the data has
missing values or unobservable (hidden) variables. To find
the parameters that can maximize the likelihood function,
EM algorithm first calculates the expectation of hidden
variables based on the current parameters estimation at the
) (8)
E-step. Then at the M-step, missing values are replaced by
the expected values calculated in the E-step to find a new
estimation of parameters that can maximize the complete
data likelihood function. These two steps are iterated to
convergence.
In the clustering applications, it is normally assumed that a
document is generated by a finite mixture model and there
is one-to-one correspondence between mixture components
and clusters. So the likelihood function p( D | θ ) , i.e. the
(9) probability of all documents D given the model parameter
θ , can be written as the equation (10):
N |C |
p ( D | θ ) = ∏∑ p(c j | θ ) p (d i | c j , θ ) (10)
i =1 j =1
where c j is j-th cluster, | C | is the number of clusters,
p(c j | θ ) is the prior distribution of
and p(d i | c j ,θ ) is the distribution of document i in cluster j.
It is further assumed that the terms are conditional
independent given the class label, and then the likelihood
function can be rewritten as the equation (11):
p ( D | θ ) = ∏ ∑ ( p (c j | θ )∏ p (t | c j , θ ))
N |C |
i =1 j =1 t∈d i
where p(t | c j , θ ) is term distribution for the term t in the
cluster j. Not all terms are equivalently relevant to the
document, so p(t | c j , θ ) can be treated as the weighted
sum of relevant distribution and irrelevant distribution as
the equation (12):
p (t | c j , θ ) = z (t ) p(t is relevant | c j , θ )
+ (1 − z (t )) p (t is irrelevant | θ )
where z (t ) = p(t is relevant ) is defined as the probability
that the term t is relevant. In addition, if the term is not
relevant, the term distribution is assumed to be the same
over different clusters and is denoted as
p(t is irrelevant | θ ) . Hence, the likelihood function can
be written as the equation (13):
N |C |
p(D | θ ) = ∏∑
i =1 j =1
( p (c j | θ )
∏ ( z (t ) p(t is relevant | c ,θ ) + (1 − z (t )) p(t is irrelevant | θ )))
i∈d i
j
To maximize this likelihood function, EM algorithm can
find a local maximum by iterating the following two steps:
(1) E-step: z) ( k +1) = E ( z | D, θ ( k ) ) (14)
)
(2) M-step: θ = arg max θ p( D | θ , z ( k ) ) (15)
) ( k +1)
)
The E-step corresponds to calculating the expected feature
relevance given the clustering result, and the M-step
corresponds to calculating a new maximum likelihood
estimation of the clustering result in the new feature space.
The proposed EM algorithm for text clustering and feature
selection is a general framework and the full
implementation is beyond the scope of this paper. Our
Iterative Feature Selection method can be fitted into this
framework. On the E-step, to approximate the expectation
of feature relevance, we use supervised feature selection
algorithm to calculate the relevance score for each term,
then the probability for the term relevance is simplified to
z ( t ) = { 0 ,1} according to whether the term relevance
score is larger than a predefined threshold value. So at each
iteration, we will remove some irrelevant terms based on
the calculated relevance of each term. On the M-step,
because K-means algorithm can be described by slightly
cluster j, extending the mathematics of the EM algorithm to the hard
threshold case (Bottou et al., 1995), we use K-means
clustering algorithm to obtain the cluster results based on
the selected terms.
4. Experiments
(11) We first conducted an ideal case experiment to demonstrate
that supervised feature selection methods, including IG and
CHI, can significantly improve the clustering performance.
Then, we evaluated the performance of unsupervised
feature selection methods, including DF, TS, TC and En in
real case. Finally we evaluated the iterative feature selection
algorithm. K-means was chosen as our basic clustering
algorithm and entropy measure and precision measure were
used to evaluate the clustering performance. Since K-means
(12) clustering algorithm is easily influenced by selection of
initial centroids, we random produced 10 sets of initial
centroids for each data set and averaged 10 times
performances as the final clustering performance. Before
performing clustering, tf*idf (with “ltc” scheme) was used
to calculate the weight of each term.
4.1 Performance Measures
Two kinds of popular measurements, entropy and precision,
were used to evaluate the clustering performance.
4.1.1 ENTROPY
Entropy measures the uniformity or purity of a cluster. Let
(13) G ', G denote the number of obtained clusters and the
number of original classes respectively. Let A denote the
set of documents in a obtained cluster, and the class label of
each document d i ∈ A, i = 1,..., | A | is denoted as label (di ) ,
which takes value c j ( j = 1..., G ) . The entropy for all
clusters is defined by the weighted sum of the entropy for
all clusters, as shown in the equation (16):
Entropy = − ∑ ∑ p jk × log( p jk )
G'
| Ak | G (16)
k =1 N j =1
where
1
Pjk = | {d i | label ( d i = c j )} | . (17)
| Ak |
4.1.2 PRECISION
Since entropy is not intuitive, we choose another measure to
evaluate the clustering performance, i.e. precision. For each
cluster, it always consists of documents from several
different classes. So we simply choose the class label which
shares with most documents in this cluster as the final class
label. Then, the precision for each cluster is defined as:
 1 Web Directory dataset, there is a significant performance
Precision( A) = max(| {d i | label ( d i ) = c j } |) (18)
| A| improvement. For example, using CHI method on Web
Directory dataset, when 98% terms are removed, the
In order to avoid the possible bias from small clusters
entropy is reduced from 2.305 to 1.870 (18.9% entropy
which have very high precision, the final precision is
reduction relatively), the precision is improved from 52.9%
defined by the weighted sum of the precision for all clusters,
to 63.0% (19.1% precision improvement relatively).
as shown in the equation (19):
Certainly, the results are just the upper bound of the
Precision = ∑
G'
| Ak | (19) clustering performance in real case because it is difficult to
Pr ecision( Ak ) select discriminative terms without prior class labels
k =1 N
information.
4.2 Data Sets
As reported in past research works (Yang & Pederen, 1997),
2.2
text categorization performance varies greatly on different
dataset. So we chose three different text datasets to evaluate 1.7
text clustering performance, including two standard labeled

Entropy
datasets: Reuters-21578 1 (Reuters), 20 Newsgroups 2 1.2
(20NG), and one web directory dataset (Web) collected
from the Open Directory Project3. There were total 21578 0.7 IG.Reuters
IG.Web
CHI.Reuters
CHI.Web
documents in Reuters, but we only chose the documents IG.20NG CHI.20NG
0.2
that have at least one topic and Lewis split assignment, and 100 40 20 10 8 6 4 2
assigned the first topic to each document before the Percentage of Selected Features
evaluation. The information about these datasets is shown
in Table 1. Figure 1. Entropy comparison on 3 datasets (supervised)
Table 1. The three datasets properties
0.81

DATA CLASSE DOCS TERMS AVG. AVG. 0.76

SETS S NUM. NUM. NUM TERMS DF PER 0.71

Precision

PER DOC TERM 0.66

0.61
REUTERS 80 10733 18484 40.7 23.6 0.56
IG.Reuters CHI.Reuters

20NG 20 18828 91652 85.3 17.5 0.51 IG.Web

WEB 35 5035 56399 131.9
4.3 Results and Analysis
4.3.1 SUPERVISED FEATURE SELECTION
First, we conducted an ideal case experiment to see whether
good terms can help text clustering. That is, we applied
supervised feature selection methods to choose the best
terms based on the class label information. Then, we
executed the text clustering task on these selected terms and
compared the clustering results with the baseline system,
which clustered the documents on full feature space.
The entropy comparison over different datasets is shown in
Figure 1, and the precision comparison is shown in Figure 2.
As can be seen, at least 90% terms can be removed with
either an improvement or no loss in clustering performance
on any dataset. With more terms removed, the clustering
performances on Reuters and 20NG changes a little, but on
1
http://www.daviddlewis.com/resources/testcollections/
2
http://www.ai.mit.edu/people/jrennie/20Newsgroups/
3 the Web Directory data, in which there are much more
http://dmoz.org/
IG.20NG
CHI.Web
CHI.20NG
0.46
11.8 100 40 20 10 8 6 4 2
Percentage of Selected Features
Figure 2. Precision comparison on 3 datasets (supervised)
Feature selection makes little progress on Reuters and
20NG, while achieves much improvement on Web
directory dataset. It motivates us to find out the reason. As
reported in the past research works on feature selection for
text categorization, the classification accuracy is almost the
same after removing some terms from Reuters and 20NG
datasets with most classifiers, including Naïve Bayesian
classifier and KNN classifier (Yang & Pedersen, 1997). We
conclude that most terms in these two datasets still have
discriminative values for classification although few terms
per class are enough to achieve acceptable classification
accuracy. In other words, there are few noisy terms in these
two datasets. Similarly, although feature selection can
reduce the dimension of the feature, the clustering
performance can not be improved significantly because
some useful terms are also ignored after removing some
terms from these two datasets. However, it is different for
noisy terms. To prove this, we conducted a Naïve Bayesian
classification experiment on Web Directory dataset and
found that the classification accuracy increased from 49.6%
to 57.6% after removing 98% terms. Hence, when these
noisy terms, such as “site, tool, category”, are removed in
clustering, the clustering performance can also be
significantly improved.
4.3.2 UNSUPERVISED FEATURE SELECTION
The second experiment we conducted is to compare the
unsupervised feature selection methods (DF, TS, TC, and
En) with supervised feature selection methods (IG and CHI).
The entropy and precision results on Reuters are shown in
Figure 3 and Figure 4.
1.3
1.25
1.2
1.15
Entropy
document frequency terms were removed by both methods,
but at the next stage, with more terms removed, the
discrimination value between the term and the class is much
more important than the document frequency for term
selection. Since discrimination value is dependent on the
class information, supervised method IG could keep those
less common but discriminative terms, such as “OPEC,
copper, drill”. However, without class information, TC
could not decide whether a term was discriminative and still
kept those common terms, such as “April, six, expect”.
Hence, it is obvious that supervised methods are much
better than unsupervised methods when more terms are
removed.
Finally, compared with other unsupervised feature selection
methods, TC is better than DF and En, but little worse than
TS. For example, as Figure 3 shows, when 96% terms are
removed, the entropies of DF and En are much higher than
the baseline value (on full feature space) but the entropies
of TC and TS are still lower than the baseline value. En is

1.1

1.05
very similar to DF because the removal of common term is
1
easily to cause more entropy reduction than the removal of
IG
DF
CHI
TC
rare term. TS is sensitive to the document similarity
0.95
TS En threshold β . When β is just suitable, the performance of
0.9
100 40 20 10 8 6 4 2 TS is better than TC. However, the threshold is difficult to
Percentage of Selected Features
tune. In addition, the time complexity of TS (at least O( N 2 ) )
Figure 3. Entropy comparison on Reuters (unsupervised) is always higher than that of TC ( O ( MN 2 ) ), because N is
0.81
always much small than N (see Table 1). Therefore, TC is
0.8
the preferred method with effective performance and low
0.79
computation cost.
0.78 Similar points can be found on the rest two datasets (20NG
Precision

0.77 and Web) (Figure 5 and Figure 6). Due to the limitations of
0.76 the paper length, we only drew the entropy measure for
0.75
IG CHI these two datasets. As can be seen from these figures, the
DF TC
0.74 TS En best performances are also yielded on Web Directory data.
0.73 About 8.5% entropy reduction and 8.1% precision
100 40 20 10 8 6 4 2 improvement are achieved while 96% features are removed
Percentage of Selected Features
with TS method.
Figure 4. Precision comparison on Reuters (unsupervised)
From these figures, we found the following points. First,
unsupervised feature selection can also improve the 2.3

clustering performance when a certain terms are removed. 2.2

For example, any unsupervised feature selection methods 2.1
can achieve about 2% entropy reduction and 1% precision
Entropy

2
improvement relatively while 90% terms are removed.
Second, unsupervised feature selection can be comparable 1.9
IG CHI
to supervised feature selection with up to 90% term 1.8 DF TC
removal. When more terms are removed, the performances 1.7
TS En

of supervised methods can be still improved, but the 100 40 20 10 8 6 4 2

performances of unsupervised methods drop quickly. In Percentage of Selected Features

order to find out the reason, we compared the terms

Figure 5. Entropy comparison on Web Directory (unsupervised)
selected by IG and TC at different removal thresholds.
After analysis, we found that at the beginning stage, low
 1.7 1.12 0.785

IG CHI
1.11
1.65
DF TC
0.78
1.1
1.6 TS En
1.09 0.775

Precision
Entropy
Entropy

1.55 1.08

1.07 0.77
1.5
IG.Entropy
1.06
1.45 CHI.Entropy 0.765
1.05 IG.Precision
CHI.Precision
1.4 1.04 0.76
100 40 20 10 8 6 4 0 1 2 3 4 5 6 7 8 9 10 11 12
Percentage of Selected Features Iteration

Figure 6. Entropy comparison on 20NG (unsupervised) Figure 7. Entropy & precision with IF selection on Reuters
2.35 0.62

4.3.3 ITERATIVE FEATURE SELECTION 2.3

0.6
2.25
The third experiment we conducted is to measure the 0.58
2.2
iterative feature selection algorithm proposed in Section 3.

Precision
Entropy
2.15 0.56
Since IG and CHI have good performance in ideal case,
2.1
they were chosen in iterative feature selection experiment. IG.Entropy 0.54
2.05
In order to speed up the iterative process, we removed the CHI.Entropy
IG.Precision 0.52
2
terms whose document frequency was lower than 3 before CHI.Precision
1.95 0.5
clustering. Then at each iteration, about 10% terms (or 3% 0 1 2 3 4 5 6 7 8 9 10 11 12
terms when less than 10% terms left) with lowest ranking Iteration
score outputted by IG or CHI were removed out. Figure 7
displays the entropy (dashed line) and precision (solid line) Figure 8. Entropy & precision with IF selection on Web Directory
results on Reuters. Figure 8 displays the results on Web
Directory dataset. We did not show the results on 20NG 12000

due to the limitations of page. 10000

Good Terms in Web
Good Terms in Reuters
As can be seen from Figure 7 and Figure 8, the performance Noisy Term in Web
Number of Terms

8000
of iterative feature selection is quite good. It is very close to Noisy Terms in Reuters

the ideal case and much better than any unsupervised 6000
feature selection methods. For example, after 11 iterations
with CHI selection on Web Directory dataset (nearly 98% 4000

terms removal), the entropy was reduced from 2.305 to

1.994 (13.5% entropy reduction relatively) and the
precision was improved from 52.9% to 60.6% (14.6%
precision improvement relatively). It is close to the upper
bound from the ideal case (18.9% entropy reduction and
19.1% precision improvement, see Section 4.3.1).
In order to verify that iterative feature selection works, we
traced the terms which were filtered out at each iteration.
Firstly, we assumed that the top 2% terms ranked by CHI in
ideal case experiment were good terms, and others were
noisy terms. Then we calculated how many “good” terms
were kept at each iteration. As can be seen from Figure 9,
the “good” terms (two bottom lines) are almost kept and
most noisy terms (two top lines) are filtered out after 10
iterations.
2000
0
1 2 3 4 5 6 7 8 9 10 11 12
Iteration
Figure 9. Feature Tracking on Reuters & Web Directory
5. Conclusions
In this paper, we first demonstrated that feature selection
can improve the text clustering efficiency and performance
in ideal case, in which features are selected based on class
information. But in real case the class information is
unknown, so only unsupervised feature selection can be
exploited. In many cases, unsupervised feature selection are
much worse than supervised feature selection, not only less
terms they can remove, but also much worse clustering
performance they yield. In order to utilize the efficient
supervised methods, we proposed an iterative feature
selection method that iteratively performs clustering and
feature selection in a unified framework. It is found that its
 performance is close to the ideal case and much better than
any unsupervised methods. Another work we have done is a
comparative study on several unsupervised feature selection
methods, including DF, TS, En, and a new proposed
method TC. It is found that TS and TC are better than DF
and En. Since TS has high computation complexity and is
difficult to tune its parameters, TC is the preferred
unsupervised feature selection method for text clustering.
References
Aggrawal, C.C., & Yu, P.S. (2000). Finding generalized
projected clusters in high dimensional spaces. Proc. of
SIGMOD’00 (pp. 70-81).
Bekkerman, R., El-Yaniv, R., Tishby, N., & Winter, Y.
(2001). On Feature Distributional Clustering for Text
Categorization. Proc. of SIGIR’01 (pp. 146-153).
Blum, A. L., & Langley, P. (1997). Selection of relevant
features and examples in machine learning. Artificial
Intelligence, 1(2), 245-271.
Bottou L., & Bengio Y. (1995). Convergence properties of
the k-means algorithms. Advances in Neural Information
Processing Systems, 7, 585-592.
Cutting, D. R., Karger, D. R., Pedersen, J. O., & Tukey.W.
(1992). Scatter/Gather: A Cluster-based Approach to
Browsing Large Document Collections. Proc. of
SIGIR’92 (pp. 318–329).
Dash, M., & Liu, H. (1997). Feature selection for
classification. International Journal of Intelligent Data
Analysis, 1(3), 131-156.
Dash, M., & Liu, H. (2000). Feature Selection for
Clustering. Proc. of PAKDD-00 (pp. 110-121).
Dempster, A. P., Laird, N. M., & Rubin, D. B. (1977).
Maximum likelihood from incomplete data via the EM
algorithm. Journal of the Royal Stat. Society,39, 1-38.
Friedman, J.H. (1987). Exploratory projection pursuit.
Journal of American Stat. Association, 82, 249-266.
Galavotti, L., Sebastiani, F., & Simi, M. (2000). Feature
selection and negative evidence in automated text
categorization. Proc. of KDD-00.
View publication stats
Jain, A.K., Duin P.W., & Jianchang, M. (2000). Statistical
pattern recognition: a review. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 22, 4-37.
Jolliffe, I.T. (1986). Principal Component Analysis.
Springer Series in Statistics.
Koller, D., & Sahami, M. (1996). Toward Optimal Feature
Selection. Proc. of ICML’96 (pp.284-292).
Koller, D., & Sahami, M. (1997). Hierarchically classifying
documents using very few words. Proc. of ICML-97 (pp.
170-178).
Kowalski, G. (1997). Information Retrieval Systems Theory
and Implementation. Kluwer Academic Publishers.
Martin, H. C. L., Mario, A. T. F., & Jain, A.K (2002).
Feature Saliency in unsupervised learning(Technical
Report 2002). Michigan State University.
Salton, G. (1989). Automatic Text Processing: The
Transformation, Analysis, and Retrieval of Information
by Computer. Addison-wesley, Reading, Pennsylvania.
Slonim, N., & Tishby, N. (2000). Document clustering
using word clusters via the information bottleneck method.
Proc. of SIGIR’00 (pp. 208-215).
Wilbur, J.W., & Sirotkin, K. (1992). The automatic
identification of stop words. Journal of Information
Science, 18, 45-55.
Wyse, N., Dubes, R., & Jain, A.K. (1980). A critical
evaluation of intrinsic dimensionality algorithms. Pattern
Recognition in Practice (pp. 415-425). North-Holland.
Yang, Y. (1995). Noise reduction in a statistical approach
to text categorization. Proc. of SIGIR’95 (pp. 256-263).
Yang, Y., & Pedersen, J. O. (1997). A comparative study
on feature selection in text categorization. Proc. of ICML-
97 (pp. 412-420).
Zamir, O., Etzioni, O., Madani, O., & Karp, R. M. (1997).
Fast and Intuitive Clustering of Web Documents. Proc. of
KDD-97 (pp. 287-290).